The detection of out-of-distribution (OOD) instances from in-distribution samples (ID) has recently garnered significant attention due to its role in handling uncertainties within data. However, current efforts primarily focus on distinguishing feature representations between OOD and ID, encountering challenges such as the inability to preserve complex geometric structures and a lack of inherent robustness to data noise. To overcome these challenges, we introduce a novel subManifold based Out-Of-Distribution detection (sMOOD) method. Our approach involves three steps, i.e., categorizing ID data into submanifolds, modeling their belonging components, and classifying new instances as OOD if the distance is considerably far from each submanifold distribution. Importantly, a theoretical analysis is provided to clarify the properties of sMOOD, establishing connections or extensions to existing methods. To validate our approach, we establish a comprehensive benchmark comparing it with various state-of-the-art methods. The experimental results demonstrate the superiority of our approach across different datasets, with both the low and full resource training scenarios.